Front Cover
Gresë Berisha
Moritz Baur
Kevin Dietrich

2025-03-17

Problem and Motivation

✈ Problem

  • Flight delays are disruptive and costly.
  • Passengers lose time ⏱ and confidence.
  • Airlines and airports face financial losses 💼 and reduced operational efficiency.

📈 Motivation

  • Tunisair aims to implement a predictive solution to anticipate delays and mitigate their impact.

🎯 Project Objective

🥅 Goal

Use machine learning to predict the length of flight delays (in minutes).

💥 Impact

  • ✔ Better scheduling
  • ✔ Reduced operational inefficiencies
  • ✔ Improved passenger satisfaction

Dataset and Evaluation

Data Source

📁 Flight data provided by Zindi, consisting of a train/test format for model development

Prediction Target:

🕒 Delay duration in minutes

Performance Metric:

📉 Root Mean Square Error (RMSE)

Exploratory Data Analysis (EDA)

🔍 Initial Insights from EDA

We analysed how delays are distributed based on:

  • Departure airports
  • Arrival airports
  • Temporal trends across the years 2016–2018

Column Description

Column Description
ID Unique flight identifier
DATOP Date of flight
FLTID Flight number
DEPSTN Departure point
ARRSTN Arrival point
STD Scheduled time of departure
STA Scheduled time of arrival
STATUS Flight status
AC Aircraft code
target Flight delay (min)

Mean Delay per Airport

by Departure and Arrival

Mean Delay per Airport

Deriving flight time from STD and STA

Removing Service Flights, i.e. flights where departure and arrival airports are the same …

Flight Time to Delay ⇒ Flight Time to Delay — After Trimming

Dissecting DATOP into YEAR, month and day of the week

Month to Sum of Flights for 2016 Month to Sum of Flights for 2017 Month to Sum of Flights for 2018 Weekday to Delay by Year

Baseline Model

Approach

A simple linear regression model using only the day of the week
(or aircraft code) as the predictor.

Performance

  • 📊 RMSE ≈ 114.69
  • 🤖 R_2 ≈ 3.01 %
Baseline Model Error

ML Model

Many categorical variables …

… but there is:

CatBoost

CatBoost
Machine-generated image—no animals were harmed

ML Model

Approach

A regression model using CatBoost with the following predictors: Flight Status; Aircraft Code; Departure and Arrival Point; Year, Month and Weekday of Departure Time

Performance

  • 📊 RMSE ≈ 96.14 (< 100)
  • 🤖 R_2 ≈ 30.48 %
ML Model Error

💬 Questions and Feedback

We're happy to answer your questions and look forward to your feedback.